Understanding over-squashing and bottlenecks on graphs via curvature
The paper addresses information distortion in Graph Neural Networks caused by topological bottlenecks called 'over-squashing.' It uses combinatorial edge curvature to prove that negatively curved edges cause these bottlenecks. The authors propose a new method, Stochastic Discrete Ricci Flow (SDRF), which surgically rewires these problematic edges to improve GNN accuracy while preserving the graph's structure.